Chelsea Finn, Assistant Professor of Computer Science and Electrical Engineering, Stanford UniversityTitle: How not to create a robot's mind
Abstract: Recent progress has demonstrated how robots can acquire complex manipulation skills from perceptual inputs through trial and error, particularly with the use of deep neural networks. Despite these successes, the ubiquitous paradigm in robot learning is to train robots in a single laboratory environment for a single, external reward objective. As a result, the generalization and versatility of robots across environment conditions, tasks, and objects remains a major challenge. And, unfortunately, our existing algorithms and training set-ups are not prepared to tackle such challenges, since these challenges demand large and diverse sets of experiences, akin to the experiences of humans. In this talk, I will briefly overview the standard approach taken by modern robot learning algorithms, and then discuss three questions that are core to acquiring more generalizable skills and behaviors. First, how can we enable robots to learn from broad sources of data and experience? Second, how can robots make decisions for longer horizon tasks? And, finally, what forms of supervision or guidance should we provide? I’ll conclude by discussing directions for future work.